Navigation positioning method based on interactive multiple model fast smooth switching framework

Abstract

The application discloses a navigation positioning method based on an interactive multi-model rapid smooth switching framework. The application firstly initializes an inertial navigation system, three sub-filter models and benchmark parameters of the interactive multi-model; then, based on innovation squared sum (ISR) and chi-square test, the original likelihood value is robustly modified to inhibit the interference of abnormal values on the likelihood value; then, based on sliding window trend analysis, two-stage determination is realized, the first stage strengthens the continuous optimal model, the second stage switches the failure model and selects a trend-qualified alternative model; then, according to the determination result, a transition probability correction coefficient is calculated to dynamically adjust a transition probability matrix (TPM) of the interactive multi-model; finally, the navigation result is output through multi-model weighted fusion, and combined navigation real-time iteration is realized, so as to solve the problems of traditional interactive multi-model, such as complex noise, model mismatch scene, likelihood value distortion, model switching lag and poor adaptability of the transition probability matrix.

Description

Technical Field

[0001] This invention belongs to the field of integrated navigation technology, and relates to intelligent filtering and state estimation of integrated navigation systems, specifically to a navigation and positioning method based on an interactive multi-model fast and smooth switching framework. Background Technology

[0002] SINS / GNSS strapdown inertial navigation systems and global navigation satellite systems combine the advantages of continuous output and strong anti-interference capabilities of strapdown inertial navigation systems with the high positioning accuracy of global navigation satellite systems, making them the mainstream navigation solutions for mobile vehicles such as drones, autonomous vehicles, and mobile robots. Interactive multi-model (IMM) algorithms, by fusing the outputs of multiple filter models, can adapt to navigation requirements under different operating conditions and are one of the core fusion algorithms of integrated navigation systems. However, traditional IMM algorithms have technical shortcomings in complex navigation scenarios: First, the transition probability matrix is ​​mostly a fixed value, unable to be dynamically adjusted according to the real-time performance of each filter model. When the current optimal model fails, the system struggles to quickly switch to a more robust alternative model, exhibiting significant model switching lag. Second, the model likelihood calculation is based on the Gaussian distribution assumption, while actual global navigation satellite system measurements are often affected by non-Gaussian noise and measurement anomalies, leading to distorted likelihood values ​​and subsequent incorrect model probability updates. Finally, traditional IMM algorithms do not incorporate changes in model probability trends, relying solely on single-point probability values ​​to update model weights, making it prone to selecting abnormal models and resulting in insufficient system robustness. Existing improvement methods mostly focus on parameter optimization of a single filter model or local adjustment of a fixed transition probability matrix, which is difficult to meet the requirements of integrated navigation systems for high precision and strong robustness under complex operating conditions. Summary of the Invention

[0003] Objective: To overcome the shortcomings of existing technologies, and addressing the problems of likelihood value distortion, model switching lag, and poor adaptability of transition probability matrices in traditional interactive multi-model algorithms caused by non-Gaussian noise, measurement anomalies, and model mismatch in complex scenarios, this invention provides a navigation and positioning method based on a fast and smooth switching framework for interactive multi-models. This method improves the accuracy of model probability updates, the speed of model switching, and the adaptability of the transition probability matrix through a collaborative design of likelihood value correction, two-level trend determination, and adaptive adjustment of the transition probability matrix, thus meeting the high-precision and robust navigation requirements of integrated navigation systems in complex environments.

[0004] Technical solution: To achieve the above objectives, the present invention adopts the following technical solution:

[0005] This invention provides a navigation and positioning method based on an interactive multi-model fast and smooth switching framework, the method comprising the following steps:

[0006] S1: Initialize the inertial navigation system, set the error parameters of the inertial measurement unit and inject the initial error, the reference transition probability matrix of the three filtering models and the interactive multi-model, load the trajectory data, and calculate the measurement error and information corresponding to the three filtering models;

[0007] S2: Based on the sum of squares of the new information and the chi-square test threshold, the original likelihood value is robustly corrected to obtain a corrected likelihood value that is resistant to abnormal interference.

[0008] S3: Extract the model probability sequence within the sliding window and perform a two-level decision: the first-level decision strengthens the continuously optimal model, the second-level decision switches the failed model and trends to a qualified alternative model, and if neither of the two-level decisions is triggered, the current transition probability matrix is ​​maintained.

[0009] S4: Calculate the transition probability correction coefficient based on the two-stage decision results. The transition probability matrix is ​​dynamically adjusted.

[0010] S5: Update the model probability based on the corrected likelihood value, obtain the state and covariance fusion result of the three models through weighted fusion, and feed the fused error estimate back to correct the inertial navigation system.

[0011] Furthermore, step S1 specifically includes the following sub-steps:

[0012] S1-1 initializes the inertial navigation system, including a standard filter model, an adaptive filter model, and a robust filter model. The state vectors of all three filters are 15-dimensional navigation error parameters. ,in , , , , These represent position error, velocity error, attitude error angle, gyroscope zero bias error, and accelerometer zero bias error, respectively. The superscript T indicates transpose.

[0013] S1-2 Set the baseline transition probability matrix in It is the baseline transition probability matrix. Refers to the model To model The transition probabilities are calculated. This matrix is ​​a 3×3 square matrix used to describe the transition probability relationship between the three filtering models. Then, the sliding window length and the two-level decision threshold are initialized to provide a basis for subsequent model probability trend analysis and decision-making.

[0014] S1-3 Calculate the new information for each model and the new covariance matrix The specific formula is as follows in, Let j be the innovation vector of the j-th model at time k. Let be the measurement vector of the j-th model at time k. Let be the observation matrix of the j-th model at time k. Let j be the one-step state prediction of the j-th model at time k. The innovation covariance matrix of the j-th model at time k, Let be the predicted state covariance matrix of the j-th model at time k. This is the measurement noise covariance matrix.

[0015] Furthermore, step S2 involves robustly correcting the original likelihood value based on the sum of squared innovations and the chi-square test, thereby suppressing the interference of non-Gaussian noise and measurement anomalies on the likelihood value calculation and improving the accuracy of model probability updates. Specifically, this includes the following sub-steps:

[0016] S2-1: Calculating the original likelihood value based on Gaussian distribution characteristics The formula is: in Let represent the likelihood value of the j-th model at time k. To predict the state covariance matrix, Let be the information covariance matrix of the j-th model at time k. To measure dimension, Let det(⋅) be the innovation vector of the j-th model at time k, det(⋅) denotes the determinant of the matrix, the superscript T is the transpose, and the superscript −1 is the inverse of the matrix. Pi is a mathematical constant, usually taken as 3.14.

[0017] S2-2: Calculate the sum of squared innovations and perform robustness correction on the original likelihood values, setting the chi-square test threshold. Set the chi-square test threshold. This value is obtained by referring to the chi-square distribution table when the measurement has 6 degrees of freedom and a significance level of 0.05. When the sum of squares of new information is less than this threshold, it indicates that the current measurement is within the 95% confidence level and is a normal measurement; when the sum of squares of new information is greater than or equal to this threshold, it indicates that the current measurement has an anomalous probability of more than 5% and needs to be downweighted.

[0018] The specific formula for calculating the sum of squares of new ideas is: Then calculate the weight correction factor. like < This indicates that the measurement is normal, and the sum of squares and weights of the new information are normal. =1.0, fully trusting the original likelihood value; otherwise, it indicates an anomaly in the measurement, and the weight of the anomaly likelihood value is reduced through exponential decay, and set... Minimum weight protection of ≥0.1 is used to prevent the likelihood value from being excessively suppressed. Final corrected likelihood value • The corrected likelihood value can effectively suppress interference from non-Gaussian noise and measurement anomalies.

[0019] Furthermore, step S3 specifically includes the following sub-steps:

[0020] S3-1: First, perform sliding window trend extraction to obtain the model probability sequence of the most recent L steps. Then, determine the optimal model, suboptimal model, and minimum model at the initial time by sorting them, and extract the complete probability change sequence of each model within the window.

[0021] S3-2: First-level judgment: Perform trend analysis on the probability change sequence of the optimal model. If the model probability at the last moment is greater than the model probability at the first moment... And the initial probability is greater than 10 times. This indicates that the model has high reliability, triggering a first-level decision to mark the model as the target model, and then adjusting the transition probability;

[0022] S3-3: Secondary Decision: If the primary decision is not triggered and the probability of the optimal model decreases, then the secondary decision is initiated: First, select from the other two models that satisfy the condition that the probability of the last model within the window is greater than 1. The model probability at the first time step within the window is times the initial probability within the window, and the initial probability within the window is greater than... The candidate models are then selected; the probability growth increment of each candidate model is calculated, which is the model probability at the last moment of the window minus the model probability at the first moment of the window; the candidate model with the largest growth increment is selected as the target model, triggering a secondary decision to achieve model switching;

[0023] S3-4: When neither of the two-level judgments is triggered, it indicates that the current probability trends of each model are stable and there is no obvious need for optimal model matching enhancement or optimal model failure switching. In this case, the current transition probability matrix remains unchanged, and the process directly proceeds to the model probability update stage in step S5.

[0024] Furthermore, step S4 specifically includes the following steps:

[0025] S4-1: Calculate the transition probability correction coefficient The formula is: in, Let be the transition probability from target model j to source model i. This represents the corrected likelihood value for the target model. Let be the transition probability from source model i to target model j. This is the corrected likelihood value for the source model. Used to correct the initial transition probability matrix;

[0026] S4-2: Utilizing Adjusting the transition probability matrix

[0027] When the first-level or second-level criteria are met, the transition probability matrix is ​​adjusted. The specific adjustment rules are as follows: The target model determined is d. This represents the adjusted transition probability from model i to model d. This represents the adjusted transition probability from model i to model m. This represents the adjusted transition probability from model i to model n. For correction factor, Let be the transition probability from model i to model d. Let be the transition probability from model i to model m. Let be the transition probability from model i to model n;

[0028] S4-3: Normalize each row of the adjusted transition probability matrix to ensure that the sum of the transition probabilities in each row is 1. The normalization formula is as follows:

[0029] The normalized transition probability matrix satisfies non-negativity and a row sum of 1, ensuring the validity of the transition probability matrix.

[0030] Furthermore, step S5 specifically includes the following steps:

[0031] State fusion: The state estimates of the three models are fused using a probabilistic weighted summation method. The formula is as follows: in, Let be the probability of the j-th model at time k. For the state estimation of the j-th model at time k, This is the state estimate after fusion.

[0032] Covariance fusion: A state bias term is introduced during the covariance matrix fusion process to improve the accuracy of covariance estimation. The formula is as follows: To ensure the positive definiteness of the covariance matrix, the fused matrix is ​​symmetric, as shown in the formula: Finally, the navigation error parameters are extracted from the fused state estimates and fed back to correct the inertial navigation system.

[0033] Beneficial Effects: Compared with existing technologies, this invention innovatively proposes a navigation and positioning method based on an interactive multi-model fast and smooth switching framework. This method, based on an interactive multi-model approach, constructs a hybrid filtering model containing three filtering models, designs a sum-of-squares innovation weighted correction mechanism to suppress abnormal measurement interference, and achieves rapid model switching through two-level trend determination and dynamic transition probability matrix correction. The integrated navigation system based on this interactive multi-model fast and smooth switching framework will be able to achieve higher robustness in more complex noisy environments. Attached Figure Description

[0034] Figure 1 This is a flowchart of the method of the present invention;

[0035] Figure 2 This is a comparison of the velocity and position errors between the traditional interactive multi-model method and the method of this invention. Detailed Implementation

[0036] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0037] This invention provides a navigation and positioning method based on an interactive multi-model fast and smooth switching framework, such as... Figure 1 As shown, this method simultaneously constructs three filtering models, suppresses abnormal measurement interference through weighted correction of the sum of squared innovations, achieves two-level decision-making based on sliding window trend analysis, and dynamically optimizes the transition probability matrix within an interactive multi-model framework, including the following steps:

[0038] S1: Initialize the inertial navigation system, set the error parameters of the inertial measurement unit and inject the initial error, the reference transition probability matrix of the three filtering models and the interactive multi-model, load the trajectory data, and calculate the measurement error and information corresponding to the three filtering models;

[0039] S2: Based on the sum of squares of the new information and the chi-square test threshold, the original likelihood value is robustly corrected to obtain a corrected likelihood value that is resistant to abnormal interference.

[0040] S3: Extract the model probability sequence within the sliding window and perform a two-stage decision: the first stage decision strengthens the continuously optimal model, the second stage decision switches the failed model and finds a suitable alternative model; if neither stage decision is triggered, the current transition probability matrix is ​​maintained.

[0041] S4: Calculate the transition probability correction coefficient based on the two-stage decision results. The transition probability matrix is ​​dynamically adjusted.

[0042] S5: Based on the corrected likelihood value, update the probabilities of the three filtering models, obtain the state and covariance fusion result of the three models through weighted fusion, and feed the fused error estimate back to correct the inertial navigation system.

[0043] In this embodiment, step S1 specifically includes the following sub-steps:

[0044] S1-1 initializes the inertial navigation system, including a standard filtering model, an adaptive filtering model, and a robust filtering model. The state vector of all three filtering models is a 15-dimensional navigation error parameter. ,in , , , , These represent position error, velocity error, attitude error angle, gyroscope zero bias error, and accelerometer zero bias error, respectively. The superscript T indicates transpose.

[0045] S1-2 Set the baseline transition probability matrix in It is the baseline transition probability matrix. Refers to the model To model The transition probabilities are calculated. This matrix is ​​a 3×3 square matrix used to describe the transition probability relationship between the three filtering models. Then, the sliding window length and the two-level decision threshold are initialized to provide a basis for subsequent model probability trend analysis and decision-making.

[0046] S1-3 Calculate the new information for each model and the new covariance matrix The specific formula is as follows in, Let j be the innovation vector of the j-th model at time k. Let be the measurement vector of the j-th model at time k. Let be the observation matrix of the j-th model at time k. Let j be the one-step state prediction of the j-th model at time k. The innovation covariance matrix of the j-th model at time k, Let be the predicted state covariance matrix of the j-th model at time k. This is the measurement noise covariance matrix.

[0047] In this embodiment, step S2, which involves robustly correcting the original likelihood value based on the sum of squares of new information and the chi-square test, is specifically performed as follows:

[0048] S2-1: Calculating the original likelihood value based on Gaussian distribution characteristics The formula is: in Let represent the likelihood value of the j-th model at time k. To predict the state covariance matrix, Let be the information covariance matrix of the j-th model at time k. To measure dimension, Let det(⋅) be the innovation vector of the j-th model at time k, det(⋅) denotes the determinant of the matrix, the superscript T is the transpose, and the superscript −1 is the inverse of the matrix. Pi is a mathematical constant, usually taken as 3.14.

[0049] S2-2: Calculate the sum of squared innovations and perform robustness correction on the original likelihood values, setting the chi-square test threshold. Set the chi-square test threshold. This value is obtained by referring to the chi-square distribution table when the measurement has 6 degrees of freedom and a significance level of 0.05. When the sum of squares of new information is less than this threshold, it indicates that the current measurement is within the 95% confidence level and is a normal measurement; when the sum of squares of new information is greater than or equal to this threshold, it indicates that the current measurement has an anomalous probability of more than 5% and needs to be downweighted.

[0050] The specific formula for calculating the sum of squares of new ideas is: Then calculate the weight correction factor. like < This indicates that the measurement is normal, and the sum of squares and weights of the new information are normal. =1.0, fully trusting the original likelihood value; otherwise, it indicates an anomaly in the measurement, and the weight of the anomaly likelihood value is reduced through exponential decay, and set... Minimum weight protection of ≥0.1 is used to prevent the likelihood value from being excessively suppressed. Final corrected likelihood value • The corrected likelihood value can effectively suppress interference from non-Gaussian noise and measurement anomalies.

[0051] Preferably, step S3 involves extracting the model probability sequence within the sliding window and performing a two-stage decision: the first stage determines to strengthen the continuously optimal model, and the second stage determines to switch the failed model and find a suitable alternative model with a valid trend. If neither stage of decision is triggered, the current transition probability matrix is ​​maintained. The specific method is as follows:

[0052] S3-1: First, perform sliding window trend extraction to obtain the model probability sequence of the most recent L steps. Then, determine the optimal model, suboptimal model, and minimum model at the initial time by sorting them, and extract the complete probability change sequence of each model within the window.

[0053] S3-2: First-level judgment: Perform trend analysis on the probability change sequence of the optimal model. If the model probability at the last moment is greater than the model probability at the first moment... And the initial probability is greater than 10 times. This indicates that the model has high reliability, triggering a first-level decision to mark the model as the target model, and then adjusting the transition probability;

[0054] S3-3: Secondary Decision: If the primary decision is not triggered and the probability of the optimal model decreases, then the secondary decision is initiated: First, select from the other two models that satisfy the condition that the probability of the last model within the window is greater than 1. The model probability at the first time step within the window is times the initial probability within the window, and the initial probability within the window is greater than... The candidate models are then selected; the probability growth increment of each candidate model is calculated, which is the model probability at the last moment of the window minus the model probability at the first moment of the window; the candidate model with the largest growth increment is selected as the target model, triggering a secondary decision to achieve model switching;

[0055] S3-4: When neither of the two-level judgments is triggered, it indicates that the current probability trends of each model are stable and there is no obvious need for optimal model matching enhancement or optimal model failure switching. In this case, the current transition probability matrix remains unchanged, and the process directly proceeds to the model probability update stage in step S5.

[0056] Preferably, step S4 involves calculating the transition probability correction coefficient based on the two-stage determination results. The specific method for dynamically adjusting the transition probability matrix is ​​as follows:

[0057] S4-1: Calculate the transition probability correction coefficient k using the following formula: in, Let be the transition probability from target model j to source model i. This represents the corrected likelihood value for the target model. Let be the transition probability from source model i to target model j. This is the corrected likelihood value for the source model. Used to correct the initial transition probability matrix;

[0058] S4-2: Utilizing Adjusting the transition probability matrix When the first-level or second-level criteria are met, the transition probability matrix is ​​adjusted. The specific adjustment rules are as follows: The target model determined is d. This represents the adjusted transition probability from model i to model d. This represents the adjusted transition probability from model i to model m. This represents the adjusted transition probability from model i to model n. For correction factor, Let be the transition probability from model i to model d. Let be the transition probability from model i to model m. Let n be the transition probability from model i to model n;

[0059] S4-3: Normalize each row of the adjusted transition probability matrix to ensure that the sum of the transition probabilities in each row is 1. The normalization formula is as follows: The normalized transition probability matrix satisfies non-negativity and a row sum of 1, ensuring the validity of the transition probability matrix.

[0060] As a preferred method, step S5 obtains the state and covariance fusion results of the three models through weighted fusion, specifically as follows:

[0061] State fusion: The state estimates of the three models are fused using a probabilistic weighted summation method. The formula is as follows: in Let be the probability of the j-th model at time k. For the state estimation of the j-th model at time k, This is the state estimate after fusion.

[0062] Covariance fusion: A state bias term is introduced during the covariance matrix fusion process to improve the accuracy of covariance estimation. The formula is as follows: To ensure the positive definiteness of the covariance matrix, the fused matrix is ​​symmetric, as shown in the formula:

[0063] Finally, the navigation error parameters are extracted from the fused state estimates and fed back to correct the inertial navigation system.

[0064] To verify the navigation and positioning method based on the interactive multi-model fast and smooth switching framework proposed in this invention, a simulation experiment was conducted using a SINS / GNSS integrated navigation system. The experiment lasted 600 seconds, with noise anomalies occurring between 120 and 200 seconds. The experimental results comparing the traditional interactive multi-model method and the method of this invention are as follows: Figure 2 As shown, from Figure 2 It can be seen that the present invention improves speed and position estimation performance compared to traditional interactive multi-model methods, especially in noisy regions. Therefore, in practical integrated navigation applications, the method of the present invention suppresses abnormal interference through innovation squared sum weighted correction and accelerates model transfer through two-level decision-making and adaptive transfer probability matrix, thereby improving the robustness of the integrated navigation system in complex environments.

Claims

1. A navigation and positioning method based on an interactive multi-model fast and smooth switching framework, characterized in that, This method simultaneously constructs three filtering models: a standard filtering model, an adaptive filtering model, and a robust filtering model. It suppresses abnormal measurement interference through weighted correction of the sum of squared innovations, achieves two-stage decision-making based on sliding window trend analysis, and dynamically optimizes the transition probability matrix within an interactive multi-model framework. The method includes the following steps: S1: Initialize the inertial navigation system, set the error parameters of the inertial measurement unit and inject the initial error, the reference transition probability matrix of the three filtering models and the interactive multi-model, load the trajectory data, and calculate the measurement error and information corresponding to the three filtering models; S2: Based on the sum of squares of the new information and the chi-square test threshold, the original likelihood value is robustly corrected to obtain a corrected likelihood value that is resistant to abnormal interference. S3: Extract the model probability sequence within the sliding window and perform a two-stage decision: the first stage decision strengthens the continuously optimal model, the second stage decision switches the failed model and finds a suitable alternative model; if neither stage decision is triggered, the current transition probability matrix is ​​maintained. S4: Calculate the transition probability correction coefficient based on the two-stage decision results. The transition probability matrix is ​​dynamically adjusted. S5: Based on the corrected likelihood value, update the probabilities of the three filtering models, obtain the state and covariance fusion result of the three models through weighted fusion, and feed the fused error estimate back to correct the inertial navigation system.

2. The navigation and positioning method based on an interactive multi-model fast and smooth switching framework according to claim 1, characterized in that, The specific steps of step S1 include: S1-1 initializes the inertial navigation system, including a standard filtering model, an adaptive filtering model, and a robust filtering model. The state vector of all three filtering models consists of 15-dimensional navigation error parameters. ,in , , , , These represent position error, velocity error, attitude error angle, gyroscope zero bias error, and accelerometer zero bias error, respectively. The superscript T indicates transpose. S1-2 sets the baseline transition probability matrix: in It is the baseline transition probability matrix. Refers to the model To model The transition probability matrix is ​​a 3×3 square matrix used to describe the transition probability relationship between the three filtering models. Then, the sliding window length and the two-level decision threshold are initialized to provide a basis for subsequent model probability trend analysis and decision. S1-3 Calculate the new information for each model and the new covariance matrix The specific formula is as follows in, Let j be the innovation vector of the j-th model at time k. Let be the measurement vector of the j-th model at time k. Let be the observation matrix of the j-th model at time k. Let j be the one-step state prediction of the j-th model at time k. The innovation covariance matrix of the j-th model at time k, Let be the predicted state covariance matrix of the j-th model at time k. This is the measurement noise covariance matrix.

3. The navigation and positioning method based on an interactive multi-model fast and smooth switching framework according to claim 2, characterized in that, The specific method for robustly correcting the original likelihood value based on the sum of squared innovations and the chi-square test threshold, as described in step S2, to obtain a corrected likelihood value resistant to abnormal interference is as follows: S2-1: Calculating the original likelihood value based on Gaussian distribution characteristics The formula is: in Let represent the likelihood value of the j-th model at time k. To predict the state covariance matrix, Let be the information covariance matrix of the j-th model at time k. To measure dimension, Let det(⋅) be the innovation vector of the j-th model at time k, det(⋅) denotes the determinant of the matrix, the superscript T is the transpose, and the superscript −1 is the inverse of the matrix. Pi is a mathematical constant. S2-2: Calculate the sum of squared innovations and perform robustness correction on the original likelihood values, setting the chi-square test threshold. This value is obtained by looking up the chi-square distribution table when the measurement has 6 degrees of freedom and a significance level of 0.

05. When the sum of squares of new information is less than the chi-square test threshold, it indicates that the current measurement is within the 95% confidence level and is a normal measurement. When the sum of squares of new information is greater than or equal to the chi-square test threshold, it indicates that the current measurement has an abnormality probability of more than 5% and needs to be downweighted. Calculate the value of the sum of squares of the new ideas The specific formula is: Then calculate the weight correction factor. : like < If the initial likelihood is 1, it indicates that the measurement is normal, and the ISR weight ω = 1.0, meaning the original likelihood value is fully trusted; otherwise, it indicates that the measurement is abnormal, and the weight of the abnormal likelihood value is reduced exponentially, and the initial likelihood is set to 1.

0. Minimum weight protection of ≥0.1; Final corrected likelihood value • .

4. The navigation and positioning method based on an interactive multi-model fast and smooth switching framework according to claim 3, characterized in that, Step S3 involves extracting the model probability sequence within the sliding window and performing a two-stage decision: the first stage determines to strengthen the continuously optimal model, and the second stage determines to switch the failed model and find a suitable alternative model with a valid trend. The specific steps for maintaining the current transition probability matrix when neither stage is triggered are as follows: S3-1: First, perform sliding window trend extraction to obtain the model probability sequence of the most recent L steps. Then, determine the optimal model, suboptimal model, and minimum model at the initial time by sorting them, and extract the complete probability change sequence of each model within the window. S3-2: First-level judgment: Perform trend analysis on the probability change sequence of the optimal model. If the model probability at the last moment is greater than the model probability at the first moment... And the initial probability is greater than 10 times. This indicates that the model has high reliability, triggering a first-level decision to mark the model as the target model, and then adjusting the transition probability; S3-3: Secondary Decision: If the primary decision is not triggered and the probability of the optimal model decreases, then the secondary decision is initiated: First, select from the other two models that satisfy the condition that the probability of the last model within the window is greater than 1. The model probability at the first time step within the window is times the initial probability within the window, and the initial probability within the window is greater than... The candidate models are then selected; the probability growth increment of each candidate model is calculated, which is the model probability at the last moment of the window minus the model probability at the first moment of the window; the candidate model with the largest growth increment is selected as the target model, triggering a secondary decision to achieve model switching; S3-4: When neither of the two-level judgments is triggered, it indicates that the current probability trends of each model are stable and there is no obvious need for optimal model matching enhancement or optimal model failure switching. In this case, the current transition probability matrix remains unchanged, and the process directly proceeds to the model probability update stage in step S5.

5. The navigation and positioning method based on an interactive multi-model fast and smooth switching framework according to claim 4, characterized in that, Step S4 involves calculating the transition probability correction coefficient based on the two-stage decision results. The specific method for dynamically adjusting the transition probability matrix is ​​as follows: S4-1: Calculate the transition probability correction coefficient The value, the formula is: in, Let be the transition probability from target model j to source model i. This represents the corrected likelihood value for the target model. Let be the transition probability from source model i to target model j. This is the corrected likelihood value for the source model. Used to correct the initial transition probability matrix; S4-2: Utilizing Value adjustment transition probability matrix When the first-level or second-level criteria are met, the transition probability matrix is ​​adjusted. The specific adjustment rules are as follows: The target model determined is d. This represents the adjusted transition probability from model i to model d. This represents the adjusted transition probability from model i to model m. This represents the adjusted transition probability from model i to model n. For correction factor, Let be the transition probability from model i to model d. Let be the transition probability from model i to model m. Let be the transition probability from model i to model n; S4-3: Normalize each row of the adjusted transition probability matrix to ensure that the sum of the transition probabilities in each row is 1, thus ensuring the validity of the transition probability matrix.

6. The navigation and positioning method based on an interactive multi-model fast and smooth switching framework according to claim 1, characterized in that, The specific method for updating the probabilities of the three filtering models based on the corrected likelihood value in step S5, obtaining the state and covariance fusion result of the three models through weighted fusion, and feeding back the fused error estimate to correct the inertial navigation system is as follows: First, the state estimates of the three filtering models are fused by weighted summation of model probabilities. Then, a state bias term is introduced during the fusion of the covariance matrix to improve the accuracy of the covariance estimation. Finally, to ensure the positive definiteness of the covariance matrix, the fused result is symmetric. Finally, the navigation error parameters in the fused state estimates are extracted and fed back to correct the inertial navigation system.

Images